Supplementary Materials

Cross-Topic Sentiment Analysis of Wikipedia Articles

A Comparative Study of AI Models

Authors W. Lewoniewski, M. Stróżyna, I. Czumałowska, A. Wojewoda, K. Węcel
Published April 2026

Table of Contents

Section S1

Description

This study compares various methods of verifying adherence of Wikipedia articles to the Neutral Point of View (NPOV) standard. To gauge potential bias, we apply four sentiment analysis models — two lexicon-based (TextBlob, VADER) and two transformer-based (RoBERTa, DistilBERT) — to nearly 7 million articles selected from the English Wikipedia.

The articles are pre-processed, categorized by topic and quality, according to proposed methodology. The results showed that sentiment of Wikipedia articles varies by topic and that proper model selection is crucial for accurate NPOV assessment. The paper contributes a practical framework for implementing sentiment analysis on longer texts. We also offer insights into how article quality correlates with sentiment outcomes.

Sentiment Analysis Wikipedia NPOV NLP Transformer Models Cross-Topic Analysis

Section S2

Methodology & Models

S2.1 Sentiment Analysis Models

Four distinct sentiment analysis models were applied to the English Wikipedia corpus, representing both traditional lexicon-based and modern transformer-based approaches. Additionally, a compound VADER variant (cVADER) was computed from VADER's compound score.

TextBlob

Lexicon-based model using pattern-based sentiment scoring. Provides polarity scores between −1 (negative) and +1 (positive).

Lexicon-Based

VADER

Valence Aware Dictionary and sEntiment Reasoner. Rule-based model with compound scores for overall sentiment.

Lexicon-Based

RoBERTa

Robustly Optimized BERT Pre-training Approach. Transformer-based deep learning model fine-tuned for sentiment classification.

Transformer

DistilBERT

Distilled version of BERT, offering a lighter transformer architecture while retaining strong sentiment classification performance.

Transformer

S2.2 Article Categorization

Wikipedia articles were categorized along two dimensions for analysis:

By Topic (18 categories)

automobile book building business city crypto event film human painting programming shop software taxon tvseries university videogame website

By Quality Score Range

0–10 11–20 21–30 31–40 41–50 51–60 61–70 71–80 81–90 91–100

Section S3

Sentiment by Topic

Sentiment analysis results for Wikipedia articles from different topics, using various models. Each bar shows the share of positive, neutral, and negative polarity.

Section S4

Sentiment by Quality Score

How sentiment distributions correlate with article quality scores (0–100 range), revealing patterns across quality tiers.

Section S5

Polarity Heatmaps

Share of Wikipedia articles with positive, neutral, and negative polarity score across different topics and quality ranges using various models.

Positive Polarity

Neutral Polarity

Negative Polarity

Section S6

Dataset & Resources

Published Dataset

The prepared dataset along with sentiment assessment by all four models is publicly available on HuggingFace. The dataset contains sentiment scores for nearly 7 million English Wikipedia articles, categorized by topic and quality.

Wikipedia Sentiment Dataset

lewoniewski/wikipedia-sentiment on HuggingFace

HuggingFace

Dataset Overview

Feature Details
Source English Wikipedia
Articles ~7,000,000
Lexicon-Based Models TextBlob, VADER (+ cVADER compound)
Transformer Models RoBERTa, DistilBERT
Sentiment Classes Positive (pos), Neutral (neu), Negative (neg)